Road Accident Severity Prediction based on Contributing Factors using Deep Learning Techniques

Abstract

Road accidents have recently emerged as the ninth most serious threat in terms of fatality causes to both individuals and governments. The cost of traffic related deaths and property damage is disproportionately high in developing countries. Investigating the elements that cause accidents and forecasting the severity of the accident is thus extremely beneficial in reducing future accidents. Existing methods for forecasting the severity of road accidents rely on shallow severity prediction models and statistical models, with only a few researchers using deep learning models. This study proposed an LSTM-based model for predicting road accidents, which forecasts indications of the severity of road accidents by learning data about road accidents. The existing studies have significant shortcomings, including the use of small datasets with limited coverage, reliance on a large set of data, low accuracy, and inability to be used in real-time. To address these challenges, we independently apply the LSTM model, CatBoost-based model and CNN model to the Balanced and Unbalanced US Accident datasets. To enhance the performance of these models, we employ three feature selection algorithms—Random Forest, Decision Tree, and Cat Boost—to extract the most relevant features from the datasets. Subsequently, the effectiveness of these various proposed methods is compared. Cat Boost combined with an LSTM model showcases the synergy between Cat Boost feature selection and LSTM modeling, based on the experimental results, it is clear that this combined approach outperforms the standalone LSTM model and the Cat Boost model, with an accuracy of 98.57.